My 3060 12gb speeds SD3.5 | Flux | SDXL | HunYuan Video | LTXV
I'm just putting my own inference speed results of my 3060 12gb here (running Windows 10). Might help someone 👍.
I might irregularly add data here but so far this is what you get (VRAM Idle usage is probably inaccurate)
To see LTXV and Hunyuan Video speeds see https://civitai.com/articles/10673
SDXL FP16 1024x1024 (Forge/Comfy)
Euler (a) normal/Karras: 1.60 it/s
DPM++ 2/3M/2s/a and all the other stuff maybe soon™️but presumably they are just slower by +-.08 s except adaptive
NOTE: SD 3.5 L/M + Flux were tested using T5xxl v1.1 Q8 GGUF and normal CLIP G/L with the SD 3.5 ComfyUI example workflow. The Clips were always able to be offloaded to RAM, my system has 24gb of RAM (16gb might get tight but idk). With 3.5 Large the Unet has to be unloaded to ram and swap in the Clips from ram and then switch them out again after CLIP text encode.
edit: i think the vram usage isnt just the unet always; until i figure that out please ignore the vram usage lol I will retest at some point with "Force/Set CLIP Device" node set to CPU, the values for Q8 SD3.5M should be correct.
Flux1 Dev Pixelwave Q6K 1024x1024 (Comfy)
Euler simple/beta/SGM Uniform: 4.49 s/it
Flux1 Dev Pixelwave Q4KM 1024x1024 (Comfy)
Euler simple/beta/SGM Uniform: 4.14 s/it
SD 3.5 Large Q8 1024x1024 (Comfy) VRAM Idle: 6.4gb | Gen: 8.9gb
Euler normal/beta/SGM Uniform: 2.82 s/it
DPM++ 2M beta: 2.88 s/it
SD 3.5 Medium FP16 1024x1024 (Comfy) VRAM Idle: 5.1gb | Gen: 5.8gb
Euler beta/SGM Uniform: fast at first, slows down ~1.03 s/it
DPM++ 2M beta/SGM Uniform: steady 1.04 s/it
SD 3.5 Medium FP16 1440x1440 (Comfy) VRAM Idle: 5.1gb | Gen: 6.8gb
Euler SGM Uniform: fast at first, slows down ~2.55 s/it
DPM++ 2M beta/SGM Uniform: steady 2.60 s/it
SD 3.5 Medium Q8 1024x1024 (Comfy) VRAM Idle: 3.2gb | Gen: 4.1gb
Euler beta/SGM Uniform: fast at first, slows down ~1.05 s/it
DPM++ 2M beta/SGM Uniform: steady 1.07 s/it
SD 3.5 Medium Q8 1440x1440 (Comfy) VRAM Idle: 3.2gb | Gen 4.9gb
Euler SGM Uniform: fast at first, slows down ~2.59 s/it
DPM++ 2M beta/SGM Uniform: steady 2.65 s/it
Extra (SD 3.5 Large Q4_1/Q5_1 3060 6gb laptop):
To see the data please see this comment made by llmaps. Big thanks to llmaps for the data
Also don't forget that since it's a laptop it will likely differ more in numbers across other laptops.
Links to some models used in this article + GGUF variant:
T5xxl v1.1 GGUF: https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf/tree/main
SD 3.5 Large + TEs/VAE: https://huggingface.co/adamo1139/stable-diffusion-3.5-large-ungated/tree/main
SD 3.5 Large GGUF: https://huggingface.co/city96/stable-diffusion-3.5-large-gguf/tree/main
SD 3.5 Medium: https://civitai.com/models/896953/stable-diffusion-35-medium
SD 3.5 Medium GGUF: https://huggingface.co/ND911/stable-diffusion-3.5-medium-GGUF/tree/main
Flux Pixelwave (Flux.1-dev 03): https://civitai.com/models/141592?modelVersionId=992642